Priori knowledge, canonical data forms, and preliminary entrentropy reduction for IVR

Abstract

Apparatus and methods for interactive voice recognition. The apparatus and methods may include a canonical phrase derivation engine configured to derive canonical phrases from voice data. The apparatus may include an input engine configured to parse utterances. The apparatus may include a knowledge extraction engine to disambiguate the utterances into words, form a sequence from the words, extract context from the sequence, pair the sequence with a phrase of the canonical phrases, merge the sequence and the phrase to form a hybrid phrase, vectorize the hybrid phrase into a vector, and feed the vector into a non-linear classification engine to determine an intent corresponding to the utterances.

What is claimed is: 
     
       1. Apparatus for interactive voice recognition, the apparatus comprising:
 a canonical phrase derivation engine configured to derive canonical phrases from voice data; 
 an input engine configured to parse utterances; 
 a knowledge extraction engine configured to:
 disambiguate the utterances into words; 
 form a sequence from the words; 
 extract context from the sequence; 
 pair the sequence with a phrase selected from a set of the canonical phrases that conforms to the context; 
 merge the sequence and the phrase to form a hybrid phrase; and 
 vectorize the hybrid phrase into a vector; and 
 
 a non-linear classification engine configured to:
 embed the vector into a classifier embedding layer; 
 feed output from the embedding layer into a bidirectional long short-term memory layer; 
 feed output from the bidirectional long short-term memory layer into a decision layer; and 
 determine an intent corresponding to the utterances. 
 
 
     
     
       2. The apparatus of  claim 1  wherein the knowledge extraction engine is configured to generate a language dimension matrix for the words. 
     
     
       3. The apparatus of  claim 2  wherein, when the language dimension matrix is a first dimension matrix, the knowledge extraction engine is further configured to generate for each of the canonical phrases a second language dimension matrix. 
     
     
       4. The apparatus of  claim 3  wherein the knowledge extraction engine is configured to:
 apply a linear model to identify the second language dimension matrix that is most similar to the first language dimension matrix; and 
 select a phrase that corresponds to a most similar second language dimension matrix. 
 
     
     
       5. The apparatus of  claim 3  wherein the knowledge extraction engine is further configured to map a word of the sequence to an element of the phrase. 
     
     
       6. The apparatus of  claim 5  wherein the knowledge extraction engine is further configured to select for the hybrid phrase, from a word of the sequence and an element of phrase, either the word or the element. 
     
     
       7. The apparatus of  claim 5  wherein the knowledge extraction engine is further configured to vectorize the hybrid phrase as input for the non-linear classification engine. 
     
     
       8. An interactive voice recognition method comprising:
 deriving canonical phrases from voice data; 
 digitally parsing utterances; 
 disambiguating the utterances into words; 
 forming a sequence from the words; 
 extracting context from the sequence; 
 pairing the sequence with a phrase selected from a set of the canonical phrases that conforms to the context; 
 merging the sequence and the phrase to form a hybrid phrase; 
 vectorizing the hybrid phrase into a vector; 
 embedding the vector into a classifier embedding layer; 
 feeding output from the embedding layer into a bidirectional long short-term memory layer; 
 feeding output from the bidirectional long short-term memory layer into a decision layer; and 
 determining an intent corresponding to the utterances. 
 
     
     
       9. The method of  claim 8  wherein the disambiguation includes forming a language-dimension matrix corresponding to the utterances. 
     
     
       10. The method of  claim 9  wherein the matrix includes a part-of-language parameter. 
     
     
       11. The method of  claim 9  wherein the matrix includes a tense parameter. 
     
     
       12. The method of  claim 9  wherein the matrix includes a coordinating term parameter. 
     
     
       13. The method of  claim 8  wherein the extracting includes a products and services tree. 
     
     
       14. The method of  claim 13  wherein the extracting includes a tie-breaking intervention by an automated attendant. 
     
     
       15. The method of  claim 8  wherein the pairing includes generating a language dimension matrix for the words. 
     
     
       16. The method of  claim 15  wherein, when the language dimension matrix is a first dimension matrix, the pairing further includes generating for each of the canonical phrases a second language dimension matrix. 
     
     
       17. The method of  claim 16  wherein;
 the pairing includes using a linear model to identify the second language dimension matrix that is most similar to the first language dimension matrix; and 
 selecting a phrase that corresponds to a most similar second language dimension matrix. 
 
     
     
       18. The method of  claim 16  wherein the merging includes mapping a word of the sequence to an element of the phrase. 
     
     
       19. The method of  claim 18  further comprising selecting for the hybrid phrase, from a word of the sequence and an element of phrase, either the word or the element. 
     
     
       20. The method of  claim 18  further comprising vectorizing the hybrid phrase.